irom77/Racegow

timed to the frame and exported as a clip with a lap-time overlay.
---
Why
Tiny-whoop racing (like the RaceGOW format) is scored on your **best 3
consecutive laps**. Finding them by hand means scrubbing through every DVR recording, frame-stepping
every gate pass, and throwing out every lap where you clipped a pole. With a dozen packs that is an
evening of work.
**Racegow does it automatically:**
| | |
|---|---|
| ⏱️ **Reads your OSD** | Finds a Betaflight mm:ss timer anywhere on screen (one is enough), reads it with template OCR, and knows exactly when you were armed. |
| 🚩 **Learns the start gate** | No track setup. It spots the parked start position in each recording, learns the lap from how the footage repeats, and pins the timing line on the start gate. Works on any layout. |
| 💥 **Throws out crashes** | Off-line laps, disarms, stalls, missed gates, laps that skip part of the course and impossibly fast "laps" are rejected, each with a reason. |
| 🛡️ **Won't fake a time** | Checks its own reference lap: if the laps it finds are too short for the track or don't chain gate to gate, you get a warning and no result, never a convincing wrong time. |
| 🏁 **Best 3 in a row** | A sliding window over every clean lap in every video, ranked globally. |
| 🎬 **Exports the highlight** | Frame-accurate cut with a live lap clock, lap 1/2/3 and total, aligned on the left edge. Rendered with HyperFrames, or pure OpenCV when Node isn't installed. |
| 🤖 **Works with your AI agent** | Ships as an Agent Skill: ask Claude Code, Codex or any coding agent *"find my best 3 laps in this folder"*. |
Real output, RaceGOW6 Track 2: 23 recordings, 72 minutes of footage, 249 laps found, 139 clean. Best: 13.85 + 13.85 + 13.73 = 41.43 s .
---
Quick start
You get fpvlaps-work/ /report.md with the result and runners-up, laps.csv with every lap,
verification images in sheets/, and the clip in export/best3_laps_ _ s.mp4.
Prefer to check before rendering? Run analyze first, look at sheets/rank1_crossings.jpg (all four
gate crossings should show the same spot), then render --rank 1.
**The official layout is the real reference.** The tool compares your laps with each other, so it
can't tell a consistently wrong line (say, a skipped dive) from the right one. Check the winning
laps against the organiser's demo flythrough. If report.md warns that the reference lap looks
wrong, re-run with --anchor VIDEO:FRAME on a start-gate crossing that begins a lap you know you
flew correctly.
**Keep one folder per track.** Put a README.md in the track folder and one subfolder per session:
analyze --videos Track2/2nd then needs no other flags. It times laps against track.json if there
is one, otherwise against the README's correct lap (which can be in another session folder), and
report.md links the demo flythrough to check against. Once a batch checks out, save-track --work
keeps its profile next to the README for later sessions. It never replaces an existing profile
without --force. Template: track-README-template.md.
Without a track README, reuse a profile with --track /track.json.
Use it from an AI coding agent
Link the skill folder where your agent looks for skills:
Then, in any repository: *"Find my fastest 3 consecutive laps in D:\FPV\Track2 and export the clip."*
The agent runs the analysis, checks the verification images, corrects the gate if needed, renders
and reports the lap table. Other agents (Cursor, Gemini CLI, Aider...) can be pointed at